Information Enginee ring (IE) Volume 3, 2014

www.se ipub.org/ie

A Study on Spectral Super-resolution of

Hyperspectral Imagery based on
Redundant Dictionary
Ying Wu*1 , Suyu Wang2, Yibin Hou3
School of Software Engineering, Beijing University of Technology, Beijing, China
wylemmon@163.com; 2suyuwang@bjut.edu.cn; 3yhou@bjut.edu.cn

*1

Abstract
With the wide application of hyperspectral imagery, the
re solution re quire d is higher and highe r. Without increasing
the equipme nt cost, a spe ctral supe r-resolution method
base d on re dundant dictionary is prese nte d in this pape r to
improve the spectral resolution of hype rspectral image ry.
The re dundant dictionary base d over-comple te signal sparse
decomposition theory is applie d to the supe r resolution of
hyperspe ctral image ry. By sparse de composing of the
spectrum curve corresponding to each pixe l along the
dire ction of spectral dime nsion, supe r-resolution is applie d
to the comple te spectral curve of each pixe l, which can
e nsure the consiste ncy of the ir spectral characteristics during
the process of supe r-resolution restoration, while the
spectral re solution of the hype rspectral image ry is improve d
e ffective ly.
Keywords
Hyperspectral Imagery; Sparse Decomposition;
Dictionary; Super-resolution Restoration

Redundant

Introduction
Hyperspectral imagery is a two-dimensional image
group obtained when the multi-spectrally imaging
spectrometer images on the same surface feature. It
comprises tens to hundreds of consecutive and
segmental spectral information, and it has a high
spectral resolution (Liguo Wang, Chunhui Zhao,
2014). With the improved spectral resolution, it is
accompanied by the increase of the cost of the
equipment when obtaining the hyperspectral imagery.
Therefore, under conditions
of constant cost, the
super-resolution restoration method which is using
the mathematical methods and signal theory to obtain
the high-resolution images has become the hot topic in
the field of improving image resolution.
The super-resolution image restoration is a method
that rebuilds a high-resolution image at the same
scene through a piece or pieces of low-resolution

images and removes noise in the original

low-resolution image (Xuefeng Yang, 2011). For
hypersepctral imagery, super-resolution algorithm is
designed for two kinds of resolutionsspatial
resolution and spectral resolution. There are extensive
studies on the spatial resolution in recent times, and
relatively less attention to spectral ones. At present,
the non-mean interpolation method is the most
intuitive way in the super-resolution restoration study
on the spectral super-resolution of hyperspectral
imagery. But due to ignoring the error introduced by
the interpolation process, the non-mean interpolation
method cannot guarantee the optimum of the entire
restoration algorithm (Xuefen Wang, Yi Yang, Jian
Cui, 2011).
In the analysis of hyperspectral imagery, Chunmei
Zhang et al (2006) noted that, in the over-complete
decomposition of the image, the design of atoms to
form a redundant dictionary should reflect the
important characteristics of the original image as
possible. Adam S. Charles (2011) establishes a
comprehensive feature information dictionary using
the method which training the actual hyperspectral
imagery pixel based on unsupervised learning to build
redundant dictionary. This method could get more
and more effective feature information. The above
results show that the sparse representation based on
redundant dictionary could describe the feature
information in hyperspectral imagery better. For this
reason, a super-resolution restoration method based
on redundant dictionary is presented in this paper. In
this method, the establishment of the high and low
resolution redundant dictionary is the core. The
method can be summarized as follows. First, establish
the training sample library with spectral curve of high
and low resolution corresponding pixel. Then, learn to
get a pair of over-complete low and high resolution
redundant dictionary by constraining the high and

67

www.se ipub.org/ie

low resolution pixels with the same coefficient. In the

restoration process, for each pixel in the hyperspectral
imagery of low resolution input, sparse decompose
the image based on low resolution redundant
dictionary to get a set of sparse coefficient firstly. Then
reconstruct the high-resolution hyperspectral imagery
using the high resolution dictionary and the obtained
sparse coefficient. Redundant dictionary application to
hyperspectral imagery spectral super-resolution
recovery can reduce the overall amount of
computation, and ensure the consistent spectral
characteristics in recovery process, effectively
improving the spectral resolution of the hyperspectral
imagery and the descriptive power of spectral
characteristics.
Spares Decomposition for Hyperspectral
Imagery
The sparse representation of signals can effectively
extract the signal characteristics. It is useful for the
subsequent processing, and also can reduce signal
processing cost in essence. The traditional signal
representation is the signal carried out based on a
group or a plurality of orthogonal basis, such as the
classical Fourier transform, wavelet transform etc. But
because the orthogonal decomposition has certain
limitations, and the composition of signal itself can be
complex, the effect will be even worse for the signal of
frequency changing in a wide range. Different from
the signal traditional orthogonal basis theory, the
over-complete sparse decomposition theory uses the
over-complete redundant function system instead of
orthogonal basis functions. The over-complete
redundant function system is also called the
redundant dictionary. It is without any restriction in
structure, and it can be composed of different basis
function. Signal sparse decomposition is a process of
founding m atoms with the optimal linear
combination to represent the signal. The method based
on redundant dictionary provides a variety of basis
function and it can adaptively select basis function
suitable for signal decomposition based on the
characteristics of the signal during the signal
expansion process, so we can get more and better
sparse representation results. And the method is more
focused on information rather than data when
describing the signal, thus, some important natural
characteristic of the original signal is often able to
capture which the traditional methods cannot get.
As a kind of the multi-dimensional data, the

68

Information Engineering (IE) Volume 3, 2014

hyperspectral imagery can also be sparse coding

through the redundant dictionary. The hyperspectral
imagery can accurately describe each pixel spectral
features using a plurality of continuous band, and the
pixel in the similar surface area has similar spectral
curve characteristic. It is resulted by the dispersion
and aliasing of the spectral curves of various types of
surface features in the area covered by the pixel. As
shown in Fig. 1, each blue line is shown as a spectral
vector of a pixel, the continuous pixel distribution
form the smooth vector curve, each red line represents
the spectrum vector of some definite kind of materials.
Although the spectral vector of pixel and endmember
shown in the picture is just the special case of
three-dimensional, in actual situation, each spectral
vector is multidimensional form in hyperplanes. But
looking the spectral characteristics as a whole, you still
can feel that it is more natural and reasonable to
sparsely decompose the pixel data.

FIG. 1 SCHEMATIC DIAGRAM OF SPECTRAL VECTOR OF PIXEL

AND ENDMEMBER

The technique of hyperspectral imagery sparse

decomposition based on redundant dictionary can
sparsely code the images from the perspective of
looking the spectral curve as a whole. This technique
can effectively maintain the spectral characteristics
consistent and accurately sparse reconstruct the
original data. Many problems faced by the
super-resolution restoration of hyperspectral imagery
are from how to ensure not to damage the spectrum
characteristics.
Hyperspectral Imagery Super-resolution
Restoration
Establish redundant dictionary for each group of
features by learning and training method. Select the
most representative spectral characteristic curve in this

Information Enginee ring (IE) Volume 3, 2014

www.se ipub.org/ie

class of objects as the atoms for the description of any

spectral curve, and to achieve efficient sparse
decomposition of the hyperspectral imagery. The
current mainstream method is to form the training
based on learning, generating atoms in dictionary
by training the data itself, for example, the best
direction method (MOD), generalized PCA method,
K-SVD method, and sparse dictionary learning
methods. In these methods, K-SVD method is the most
common method currently (Mianzheng Lu, 2012; Jie
Ren, 2013; R. Rubinstein, 2010; Rim Walha, 2013). The
K-SVD method is adopted in this paper to train the
redundant dictionary.
First, for each type of typical objects, select a group
hyperspectral imagery which contains enough spectral
information as the training sample, determining the
dictionary which can describe the spectral
characteristics effectively by learning and training
method. Take = [ 1 , 2 ] as the training sample
data, = [ 1 , 2 ] as Sparse matrix of sparse
decomposition. So the design process of the dictionary
is the process of resolving the constrained
optimization problems of the Eq. 1.
arg minD , 2F , . . , 0

where presents

(1 )

the column i of , p

presents the Frobenius norm, that is = ij 2 .

The optimization process is the process of renewal of

the coefficient and the dictionary. It means looking for
a group of the most representative spectral
characteristics on the objects to make each training
sample achieve optimal sparse decomposition based
on the dictionary and at the same time any
hyperspectral imagery obtained by the same remote
sensing could achieve good sparse decomposition
based on the dictionary. This paper uses K-SVD
algorithm to solve the optimization problem.
Based on the redundant dictionary design method
about object classes, we study a constraint learning
Input low
resolution
image

Sparse
decomposition

low resolution
dictionary

design method of the high and low resolution

redundant dictionary. First, for each type of typical
object, selecting a group of hyperspectral imagery
contains enough information of the high resolution
spectral features. Then selecting part band as the low
resolution image. Define the high resolution pixels in
the training sample as , corresponding to the low
resolution pixels as , so the corresponding
resolution dictionary can be obtained by minimizing
the following objective equation.
{ , , } = arg min , , 22 +

22 + 1

(2 )

In the Eq. 2, presents the vector composed of sparse

coefficient . and are the optimal high and low
resolution redundant dictionary obtained by training.
1 is the regularization form problem for solving
the optimization problem. In order to ensure that it
could use the decomposition coefficients of the low
resolution pixels and high resolution redundant
dictionary to rebuild the high resolution hyperspectral
imagery, we need to keep the decomposition
coefficient of high resolution pixel decomposed on
the high resolution dictionary consistent with the
decomposition coefficient of low resolution pixel
decomposed on the low resolution dictionary.
After building the high and low resolution redundant
dictionaries, for the any pixel in the low resolution
hyperspectral imagery, we could obtain the sparse
representation coefficients through the Eq. 3.
2

= arg min 2 + 1

(3)

(4)

Accordingly, the pixel of high resolution hyperspectral

image could be obtained by the Eq. 4.
Through the detailed description of the process of the
super resolution restoration on the hyperspectral
imagery spectral above, the whole process can be
simply expressed as what the Fig. 2 shows.

sparse coe fficie nt

mode l constraint

Sparse
reconstruction

Output high
resolution
image

high re solution
dictionary

FIG. 2 THE PROCESS OF SPECTRAL S UPER-RESOLUTIONG OF HYPERSPECTRAL IMAGERY BASED ON REDUNDANT DICTIONARY

69

www.se ipub.org/ie

Information Engineering (IE) Volume 3, 2014

Experimental Results
The hyperspectral imagery materials are all got from
the hyperspectral remote sensing OMIS. We select the
three pieces of hyperspectral image River1, River2,
River3 as the experimental data. Each piece is
composed of 128 bands, 512 * 512 pixels. Select River2
as the dictionary training sample set, because it
contains a greater variety of objects, and it also
contains most objects in River2 and River3. So the
information in River2 is more comprehensive and
more general.
We have two ways to evaluate the result of the
super-resolution of hyperspectral imagery in this
work. One is the objective value comparison, the other
is the subjective judgment. They all compare with the
original high-resolution hyperspectral images and the
interpolation reconstruction hyperspectral images. We
use the PSNR (Pea k Signal to Noise Ratio) as the
objective value evaluation criteria.
The experiment results show that the hyperspectral
imagery reconstructed by this method proposed in
this paper is closer to the original hyperspectral
imagery than the interpolation method, and the PSNR
value is higher than the interpolation method, the
specific value is shown in Table 1.
TABLE 1 THE PSNR VALUE COMPARE BETWEEN RECONSTRUCTION
IMAGES AND INTERPOLATION IMAGES OF RIVER1 AND RIVER3

we will research how to improve the spatial resolution

and spectral resolution of hyperspectral imagery at the
same time. The restored image obtained by the
method proposed in this paper still has a gap with the
original in some bands. In future research, we will
optimize the training process of the dictionary and the
restored map can improve image quality by some
methods, to achieve better results.
ACKNOWLEDGMENT

This work was supported by the National Natural

Science Foundation of China No.61201361, Science
Foundation of The Beijing Education Commission
No.KM201310005028,
Training
Programme
Foundation for the Talents in Beijing City
No.2013D005015000008,
Training
Programme
Foundation for the Youth Talents by Beijing Education
Commission. We also thank the anonymous reviewers
for their insightful comments.
REFERENCES

Adam S. Charles, Bruno A. Olshause n, and Christophe r J.

Roze ll. Learning Sparse Codes for Hyperspectral
Image ry. IEEE Journal of Se lecte d Topics in Signal
Processing, 2011, 5(5): 963-978.
Jie Re n, Jiaying Liu. Conte xt-Aware Spares Decomposition
for Image De noising and Supe r-Resolution. IEEE

Number

River1
reconstr
uction

River1
interpola
tion

River3
reconstr
uction

River3
interpola
tion

Whole
band

Liguo Wang, Chunhui Zhao. Processing Te chniques of

38.242919

33.478417

41.651154

32.185499

Hype rspe ctral Image ry. National De fe nse Industry

transactions on image processing, 22(4), 2013: 199-208.

Press. Be ijing, 2013, 5.

It is obviously that the data in Table 1 shown this

method improving the spectral resolution of
hyperspectral imagery more effectively.
Conclusions
The experimental results show that the super
resolution restoration method proposed in this paper
can effectively improve the spectral resolution of
hyperspectral imagery. According to the information
collected from the same remote sensor at the same
area, the high spectral resolution hyperspectral
imagery can be reconstructed by the low spectral
resolution hyperspectral imagery, which could reduce
the equipment cost required in the acquisition of
hyperspectral imagery.
The successful application of this method has laid a
good foundation for the later research. Furthermore,

70

Mianzhe ng Lu, Qihe ng Zhang, Zhiyong Xu, Zhe nming

Pe ng. Ove r-complete sparse re prese ntation of the image
supe r-resolution

reconstruction

me thod.

Systems

Enginee ring and Electronics, 34(2), 2012:42-60.

R. Rubinste in, A. M. Bruckste in, M. Elad. Double Sparsity:
Learning

Sparse

Approximation.

Dictionaries

for

Sparse

Signal

IEEE Tans. Signal. Process. 9, 2010:

213-287.
Rim Walha, Fadoua Drira, Franck Le bourge ois, Christophe
Garcia, Ade l M. Alimi. Multiple Learne d Dictionaries
base d Clustere d Sparse Coding for the Supe r-Resolution
of Single Te xt Image, Inte rnational Confe rence on
Docume nt Analysis and Recognition. 12, 2013.
Xue fe n Wang, Yi Yang, Jian Cui. The Study on Image
Super-Resolution Reconstruction Processing Algorithm.

Information Enginee ring (IE) Volume 3, 2014

Lase r & Infrare d, 11(41), 2011: 70-85.

Xue fe ng

Yang.

Research

on

Supe r-Resolution

Reconstruction Technology in Fre que ncy and Spatial.

Domain for Re mote Se nsing Images (2011).
Wu Ying was born in Be ijing, China in 1989. She rece ive d
the B.E. de gree in software e ngineering from China
Unive rsity of Ge oscie nces in 2007 and the M.E. de gree in
software e ngineering from Be ijing University of
Technology in 2014, Be ijing, China.
She studie d in school of software e nginee ring of Be ijing
Unive rsity of Te chnology. He r research inte rests include
image and vide o processing, sparse re prese ntation and
compressive se nsing.
Wang Suyu re ce ive d the B.E. de gree in He be i Unive rsity,
He be i, China in 1999, the M.E. de gree in Sichuan Unive rsity,
Sichuan, China in 2002, the Sc.D. degree in Be ijing
Unive rsity of Technology, Be ijing, China in 2008, and has the
post-doctoral research in Be ijing Unive rsity of Technology
from 2008 to 2010. She has joint research and unde rtake n the
Research Associate work in de partme nt of e le ctronic and

www.se ipub.org/ie

information
Unive rsity.

e nginee ring

of

Hong

Kong

Polytechnic

She participate d in more than 10 research projects, publishe d

more than 30 pape rs, participate d in two patents and 4
software copyrights. Her research inte rests include image
and video processing, e mbe dde d syste ms and so on.
Hou Yibin re ce ive d the M.E. de gree in de partme nt of
computer scie nce from Xi'an Jiaotong University, Xian,
China, the Sc.D. degree in de partme nt of e lectrical
e ngineering from EINDHOVEN Unive rsity of Technology,
Holland.
He was a professor of computer scie nce , compute r scie nce
doctoral tutor, vice preside nt of e nginee ring studies, director
of the compute r institute in Xian Jiaotong Unive rsity from
1990 to 2002, dean of School of Software in Be ijing
Unive rsity of Technology , director of e mbe dde d computing
institute , dire ctor of software and ne tworking systems
e ngineering technology research cente r, Be ijing since 2002 .
He publishe d more than 100 papers, a scholarly monograph,
5 inve ntion pate nts and a scie nce film. He r research interests
include software e ngineering, embe dde d computing,
ne tworking te chnology and applications.

71